Re: [RFC PATCH] Bluetooth: hci_event: Ignore multiple conn complete events

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



Hi Sönke,

On Fri, Jan 21, 2022 at 3:52 PM Sönke Huster <soenke.huster@xxxxxxxxx> wrote:
>
> Hi Luiz,
>
> On 22.01.22 00:32, Luiz Augusto von Dentz wrote:
> > Hi Sönke,
> >
> > On Fri, Jan 21, 2022 at 3:18 PM Sönke Huster <soenke.huster@xxxxxxxxx> wrote:
> >>
> >> Hi Luiz,
> >>
> >> On 21.01.22 22:31, Luiz Augusto von Dentz wrote:
> >>> Hi Sönke,
> >>>
> >>> On Fri, Jan 21, 2022 at 10:22 AM Sönke Huster <soenke.huster@xxxxxxxxx> wrote:
> >>>>
> >>>> I just noticed that just checking for handle does not work, as obviously 0x0 could also be a handle value and therefore it can't be distinguished, whether it is not set yet or it is 0x0.
> >>>
> >>> Yep, we should probably check its state, check for state != BT_OPEN
> >>> since that is what hci_conn_add initialize the state.
> >>>
> >>
> >> I thought there are more valid connection states for the first HCI_CONNECTION_COMPLETE event, as it also occurs e.g. after an HCI_Create_Connection command, see Core 5.3 p.2170:
> >>> This event also indicates to the Host which issued the HCI_Create_Connection, HCI_Accept_-
> >>> Connection_Request, or HCI_Reject_Connection_Request command, and
> >>> then received an HCI_Command_Status event, if the issued command failed or
> >>> was successful.
> >>
> >> For example in hci_conn.c hci_acl_create_connection (which triggers a HCI_Create_Connection command as far as I understand), the state of the connection is changed to BT_CONNECT or BT_CONNECT2.
> >> But as I am quite new in the (Linux) Bluetooth world, I might have a wrong understanding of that.
> >
> > Yep, we would probably need a switch to capture which states are valid
> > and which are not or we initialize the handle with something outside
> > of the valid range of handles (0x0000 to 0x0EFF) so we can initialize
> > it to e.g. 0xffff (using something like define HCI_CONN_HANDLE_UNSET)
> > so we can really tell when it has been set or not.
> >
>
> I think the state switch is just possible if there is no possibility
> to change a connection state back into one of the accepted states.
> Unless changing the state back into an accepted state includes a call
> to "hci_conn_del_sysfs", as the real issue when getting a duplicate
> HCI_Create_Connection event is that device_add in hci_conn_add_sysfs
> is called twice for the same connection.
>
> There might be other issues as well in processing a duplicate event,
> but as far as I can see the bugs I trigger rely on multiple calls to
> device_add which lead in the long run to multiple user-after frees
> or null-pointer derefs. I tried to write that up in the bugzilla report
> here: https://bugzilla.kernel.org/show_bug.cgi?id=215497
>
>
> When using something like HCI_CONN_HANDLE_UNSET, we need to make sure
> that everywhere where we receive a handle from an event and use it to
> set conn->handle, it is a valid one. Otherwise a hacked / malicious
> controller would just send multiple events for the invalid handle.

Well that is already the case for 0x0000, so one can in fact already
abuse such a thing, so yes we probably should check if the received
handle is within the valid range or not.

> What solution do you prefer? If you don't mind I'd like to try to
> create a patch.

I'd start by initializing the conn->handle to 0xffff and add checks
for it later we can add checks for received handle as well.

> >>>> On 21.01.22 18:36, Soenke Huster wrote:
> >>>>> When a HCI_CONNECTION_COMPLETE event is received multiple times
> >>>>> for the same handle, the device is registered multiple times which leads
> >>>>> to memory corruptions. Therefore, consequent events for a single
> >>>>> connection are ignored.
> >>>>>
> >>>>> The conn->state can hold different values so conn->handle is
> >>>>> checked to detect whether a connection is already set up.
> >>>>>
> >>>>> Buglink: https://bugzilla.kernel.org/show_bug.cgi?id=215497
> >>>>> Signed-off-by: Soenke Huster <soenke.huster@xxxxxxxxx>
> >>>>> ---
> >>>>> This fixes the referenced bug and several use-after-free issues I discovered.
> >>>>> I tagged it as RFC, as I am not 100% sure if checking the existence of the
> >>>>> handle is the correct approach, but to the best of my knowledge it must be
> >>>>> set for the first time in this function for valid connections of this event,
> >>>>> therefore it should be fine.
> >>>>>
> >>>>> net/bluetooth/hci_event.c | 11 +++++++++++
> >>>>>  1 file changed, 11 insertions(+)
> >>>>>
> >>>>> diff --git a/net/bluetooth/hci_event.c b/net/bluetooth/hci_event.c
> >>>>> index 681c623aa380..71ccb12c928d 100644
> >>>>> --- a/net/bluetooth/hci_event.c
> >>>>> +++ b/net/bluetooth/hci_event.c
> >>>>> @@ -3106,6 +3106,17 @@ static void hci_conn_complete_evt(struct hci_dev *hdev, void *data,
> >>>>>               }
> >>>>>       }
> >>>>>
> >>>>> +     /* The HCI_Connection_Complete event is only sent once per connection.
> >>>>> +      * Processing it more than once per connection can corrupt kernel memory.
> >>>>> +      *
> >>>>> +      * As the connection handle is set here for the first time, it indicates
> >>>>> +      * whether the connection is already set up.
> >>>>> +      */
> >>>>> +     if (conn->handle) {
> >>>>> +             bt_dev_err(hdev, "Ignoring HCI_Connection_Complete for existing connection");
> >>>>> +             goto unlock;
> >>>>> +     }
> >>>>> +
> >>>>>       if (!ev->status) {
> >>>>>               conn->handle = __le16_to_cpu(ev->handle);
> >>>>>
> >>>
> >>>
> >>>
> >>
> >> Best
> >> Sönke
> >
> >
> >



-- 
Luiz Augusto von Dentz




[Index of Archives]     [Bluez Devel]     [Linux Wireless Networking]     [Linux Wireless Personal Area Networking]     [Linux ATH6KL]     [Linux USB Devel]     [Linux Media Drivers]     [Linux Audio Users]     [Linux Kernel]     [Linux SCSI]     [Big List of Linux Books]

  Powered by Linux